Data Processing for Effective Federated Learning in Future Networks

Abstract

Artificial intelligence (AI) for networks and networks for AI are crucial development paths
for future telecommunication networks. With data distributed across multiple network
providers, distributed AI is needed, to facilitate collaborative model training while ensuring no data leakage. By integrating distributed AI and machine learning (ML) with advanced
network technologies such as Network Function Virtualization (NFV), future network architectures can achieve enhanced performance, including improved scalability, reduced operational
costs and lower latency. As an important branch of distributed ML, federated learning (FL) offers
several advantages, including reduced data exposure, improved scalability, and minimized data
transmission. However, although FL can be adapted to network applications, the performance of
FL models is still hampered by network data quality. These data can be collected from the typical
applications of NFV architecture like the Internet of Things (IoT) and edge computing. These
network data can be also generated from virtual network functions (VNFs) which are sequentially ordered as a service function chain (SFC) to realize end-to-end (E2E) network services. The
gathered data often faces issues such as missing values, noise, and non-independent and identically distributed (non-IID) distributions. Under these conditions, the performance of FL models
will be harmed when facing network applications. As such, traditional centralized/distributed
data processing technology faces challenges when FL is applied to complex network scenarios.
This thesis focuses on how to process data to improve the performance of FL models in future
distributed networks and application scenarios.
This thesis creates a road map for addressing the data processing problem in FL under various
network scenarios. Two story lines run through the entire thesis. First, from the perspective of
the data processing timeline, the data processing before FL, during FL and after FL is studied.
Secondly, from the angle of FL types, horizontal federated learning (HFL) and federated transfer
learning (FTL) are explored, with future ongoing work on vertical federated learning (VFL).
Correspondingly, before FL, the improved one-class support vector machine method can realize
optimized feature selection for HFL in IoT networks. During FL, the proposed distributed data
augmentation and federated transfer component analysis achieve effective knowledge transfer
for FTL in VNFs. At last, after FL from different network service providers (NFPs), the proposed
Viterbi-based algorithm can quickly implement delay-constrained new SFC partitioning and crossdomain deployment, with transfer learning (TL) for VNF auto-scaling. In practical scenarios,
combining state-of-the-art FL platforms in a real network test bed bridges the gap between
theoretical and practical algorithms’ performance. The thesis conducts data processing research
on FL towards future AI-NFV network architecture, which is of good value in unleashing the
huge potential of distributed AI in future telecommunication networks.

Date of Award	17 Jun 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Dimitra Simeonidou (Supervisor) & Shadi Moazzeni (Supervisor)